What might Faceseek do under the hood to match age-old photos?

I was messing around Faceseek by feeding it a selfie from years ago. It returned a match to a grainy, old photo from 2015 that I didn’t even remember posting. The accuracy was uncanny.

I’m curious: what kind of reverse-engineering approach would you use to figure out how this thing works? Is it about neural nets, hashing, facial landmarks, or something more arcane like embedding vectors?

Would be awesome to see how someone in this community would start unraveling it—especially around real-world errors like aging, lighting shifts, or low-res inputs. Anyone tried poking at similar face matching tools?